Inference for identifiable parameters in partially identified econometric models

Size: px
Start display at page:

Download "Inference for identifiable parameters in partially identified econometric models"

Transcription

1 Journal of Statistical Planning and Inference 138 (2008) Inference for identifiable parameters in partially identified econometric models Joseph P. Romano a,b,, Azeem M. Shaikh c a Department of Economics, Stanford University, USA b Department of Statistics, Stanford University, USA c Department of Economics, University of Chicago, USA Available online 12 March 2008 Abstract This paper considers the problem of inference for partially identified econometric models. The class of models studied are defined by a population objective function Q(θ, P ) for θ Θ. The second argument indicates the dependence of the objective function on P, the distribution of the observed data. Unlike the classical extremum estimation framework, it is not assumed that Q(θ, P ) has a unique minimizer in the parameter space Θ. The goal may be either to draw inferences about some unknown point in the set of minimizers of the population objective function or to draw inferences about the set of minimizers itself. In this paper, the object of interest is some unknown point θ Θ0(P ), where Θ0(P ) = arg min θ Θ Q(θ, P ), and so we seek random sets that contain each θ Θ0(P ) with at least some prespecified probability asymptotically. We also consider situations where the object of interest is the image of some point θ Θ0(P ) under a known function. Computationally intensive, yet feasible procedures for constructing random sets satisfying the desired coverage property under weak assumptions are provided. We also provide conditions under which the confidence regions are uniformly consistent in level Elsevier B.V. All rights reserved. Keywords: Extremum estimation; Partially identified model; Incomplete model; Identified set; Identifiable parameter; Subsampling; Uniform coverage; Confidence region; Interval regression; Moment inequalities 1. Introduction This paper provides computationally intensive, yet feasible methods for inference for a large class of partially identified econometric models. A partially identified model is a model in which the parameter of interest is not uniquely defined by the distribution of the observed data. Such models arise naturally in many parts of empirical work in economics. The class of models considered is defined by a population objective function Q(θ, P ) for θ Θ. The second argument indicates the dependence of the objective function on P, the distribution of the observed data. Unlike the classical extremum estimation framework, it is not assumed that Q(θ, P ) has a unique minimizer in the parameter space Θ. The goal may be either to draw inferences about some unknown point in the set of minimizers of the population objective function or to draw inferences about the set of minimizers itself. In this paper, we consider the first of these Corresponding author at: Department of Economics, Stanford University, USA. addresses: romano@stanford.edu (J.P. Romano), amshaikh@uchicago.edu (A.M. Shaikh) /$ - see front matter 2008 Elsevier B.V. All rights reserved. doi: /j.jspi

2 J.P. Romano, A.M. Shaikh / Journal of Statistical Planning and Inference 138 (2008) two goals. The object of interest is some unknown point θ Θ 0 (P ), where Θ 0 (P ) = arg min Q(θ, P ). (1) θ Θ We henceforth refer to any θ Θ 0 (P ) as an identifiable parameter and Θ 0 (P ) as the identified set. In this instance, given i.i.d. data X i,i = 1,...,n, generated from P, we seek random sets C n = C n (X 1,...,X n ) that contain each identifiable parameter with at least some prespecified probability asymptotically. That is, for all θ Θ 0 (P ) we have lim inf P {θ C n} 1 α. (2) We term such sets as confidence regions for identifiable parameters that are pointwise consistent in level. This terminology reflects the fact that the confidence regions are valid only for a fixed probability distribution P and helps distinguish this coverage requirement from others discussed later in which we will demand that the confidence regions are valid uniformly in P. We construct random sets C n satisfying (2) under weak assumptions on P. These assumptions are formulated in terms of restrictions on the asymptotic behavior of the estimate of the population objective function, ˆQ n (θ). Most often ˆQ n (θ)=q(θ, Pˆ n ) for some estimate Pˆ n of P. Our construction is based on the usual duality between confidence regions and hypothesis testing: we invert tests of each of the null hypotheses H θ : θ Θ 0 (P ) for θ Θ that control the usual probability of a Type 1 error at level α. Typically, Θ is a subset of Euclidean space and so the number of null hypotheses may be uncountably infinite. The idea of inverting tests using computer-intensive methods can be traced to Diciccio and Romano (1990). In the second goal, the object of interest is the identified set, Θ 0 (P ), itself. In this case, given i.i.d. data X i,i=1,...,n, generated from P, we seek instead random sets C n = C n (X 1,...,X n ) that contain the identified set with at least some prespecified probability asymptotically. The problem of constructing such sets is treated in a companion paper, Romano and Shaikh (2006b). Our results on confidence regions for identifiable parameters build upon the earlier work of Chernozhukov et al. (2007), who were the first to consider inference for the same class of partially identified models. Our analysis also extends the work of Imbens and Manski (2004), who analyze the special case of the above class of partially identified models in which the identified set is an interval whose upper and lower endpoints are means or at least behave like means asymptotically. The authors discuss different definitions for confidence regions in the context of this setting and argue forcefully that it may be desirable to demand validity of confidence regions not just for each fixed probability distribution P satisfying weak assumptions, but rather uniformly over all P in some large class of distributions P. Confidence regions that fail to satisfy this requirement have the feature that for every sample size n, however large, there is some probability distribution P P for which the coverage probability of the confidence region under P is not close to the prescribed level. Researchers may therefore feel that inferences made on the basis of asymptotic approximations are more reliable if the confidence regions exhibit good uniform behavior. Of course, such a requirement will typically require restrictions on P beyond those required for pointwise consistency in level. Bahadur and Savage (1956), for example, show that if P is suitably large, then there exists no confidence interval for the mean with finite length and good uniform behavior. Romano (2004) extends this nonexistence result to a number of other problems. We provide restrictions on P under which both types of confidence regions suggested in this paper have good uniform behavior. Concretely, we provide conditions under which C n satisfies lim inf inf inf P {θ C n} 1 α. (3) θ Θ P P:θ Θ 0 (P ) By analogy with our earlier terminology, sets satisfying (3) are referred to as confidence regions for identifiable parameters that are uniformly consistent in level. Note that if the identified set Θ 0 (P ) consists of a single point θ 0 (P ), then this definition reduces to the usual definition of confidence regions that are uniformly consistent in level; that is, lim inf inf P {θ 0(P ) C n } 1 α. P P In order to provide conditions under which these coverage properties hold, we use recently developed results that provide general conditions under which confidence regions for a parameter constructed using subsampling are uniformly consistent in level. We have so far assumed that the object of interest is an identifiable parameter, θ Θ 0 (P ). More generally, the object of interest may be the image of an identifiable parameter under a known function. A typical example of such a function

3 2788 J.P. Romano, A.M. Shaikh / Journal of Statistical Planning and Inference 138 (2008) is the projection of R k onto one of the axes. We extend the above definitions of confidence regions to this setting as follows. Consider a function f : Θ Λ. Denote by Λ 0 (P ) the image of Θ 0 (P ) under f ; that is, Λ 0 (P ) ={f(θ) : θ Θ 0 (P )}. (4) We refer to a set C f n as a confidence region for a function of identifiable parameters that is pointwise consistent in level if for any λ Λ 0 (P ) we have that lim inf P {λ Cf n } 1 α. (5) As before, we may also demand uniformly good behavior over a class of probability distributions P by requiring that C f n satisfy lim inf inf inf P {λ λ Λ P P:λ Λ 0 (P ) Cf n } 1 α. (6) By analogy with our earlier terminology, sets satisfying (6) are referred to as confidence regions for a function of identifiable parameters that are uniformly consistent in level. We adapt our constructions of confidence regions for identifiable parameters to provide constructions of confidence sets satisfying these alternative coverage requirements. Methods for drawing inferences in partially identified models have allowed economists to solve many empirical problems that previously either were intractable or relied on untenable assumptions to achieve identification. Manski (2003) has argued forcefully against invoking such assumptions to make inferences in the context of missing outcome data and treatment response, as it degrades the credibility of inferences made under those assumptions. Another class of models that may only be partially identified is given by game-theoretic models with multiple equilibria. When confronted with such models, researchers often impose identification by assuming some sort of an ad hoc equilibrium selection mechanism. More recently, the empirical literature has explored inferences based on exploiting only those restrictions implied by equilibrium behavior that do not depend on the particular equilibrium being played by the agents. These restrictions are often defined by a system of moment inequalities, which is a special case of the class of models considered in this paper. Cilberto and Tamer (2004) and Borzekowski and Cohen (2005), for example, use this idea to analyze an entry model and a coordination game, respectively. Benkard et al. (2007) consider the problem of making inferences in dynamic models of imperfect competition, where in the absence of additional assumptions, the unknown parameters of the model are naturally restricted to a set defined by moment inequalities. Andrews et al. (2004), Pakes et al. (2005) and Rosen (2007) consider other economic applications of systems of moment inequalities, but develop independent methods of inference. These methods all assume some conservative upper bound on the number of moments where the inequality holds with equality and are therefore more conservative than the methods developed in this paper. Andrews and Guggenberger (2007) avoid this problem by considering the use of subsampling for models defined by moment inequalities. Andrews and Soares (2007) consider the use of moment selection techniques to select the moments for which the inequality holds with equality in a first stage and then use multivariate normal approximations to construct critical values. Similar ideas have been proposed by Bugni (2007) and Canay (2007), who use the bootstrap instead of multivariate normal approximations to construct critical values. Such methods, however, depend on the unique structure of the moment inequality problem and therefore cannot be easily generalized to more general partially identified models. Importantly, the results in this paper apply not only to systems of moment inequalities, but also to other examples of partially identified models. Bajari et al. (2007) consider inference in a semiparametric discrete choice model, which gives rise to a maximum score-type of objective function with possibly more than one maximizer. Santos (2007) considers the problem of inference in a nonparametric regression model with identification through instrumental variables. In this case, if one is not willing to assume that the instrumental variables are a complete and sufficient statistic for the regressors, then the model is only partially identified. This example is also noteworthy because the object of interest in this case, the unknown regression function, is infinite dimensional. Beresteanu and Molinari (2006) propose an alternative approach to inference in partially identified models that is based on the theory of random sets. Their approach applies to a more restrictive class of models than considered here. Moreover, their construction of confidence regions for identifiable parameters is more conservative than ours. Galichon and Henry (2006) also provide an approach to inference for partially identified models, but it too is more conservative than ours.

4 J.P. Romano, A.M. Shaikh / Journal of Statistical Planning and Inference 138 (2008) The remainder of the paper is organized as follows. We begin in Section 2 by providing a number of concrete examples that fall within the class of models we are considering. These examples not only serve as motivation, but will also be useful later for illustrating our methodology. In Section 3, we analyze the problem of constructing confidence sets that satisfy the coverage requirements (2) and (3). We then extend this methodology to construct confidence regions satisfying (5) nd (6). Finally, we conclude and provide directions for future research in Section Some motivating examples In this section, we provide motivation for our study of partially identified models by describing several specific examples of such models. We will return to these examples repeatedly in the sequel as illustrations of our methodology. Example 2.1 (One-sided mean). Perhaps the simplest example of a partially identified model is given by the following setup. Let X i,i = 1,...,n, be an i.i.d. sequence of random variables with distribution P on R. Denote by μ(p ) the mean of the distribution P. The parameter of interest, θ 0, is known to satisfy θ 0 μ(p ). For example, θ 0 might be the mean of an unobservable random variable Z i with distribution Q on R that is known to satisfy Z i X i, and so μ(q) μ(p ). The identified set is therefore given by Θ 0 (P ) ={θ R : θ μ(p )}. We may characterize this set as the set of minimizers of the population objective function Q(θ, P ) = (μ(p ) θ) 2 +, where the notation (a) + is used as shorthand for max{a,0}. The sample analog of Q(θ, P ) is given by ˆQ n (θ)=( X n θ) 2 +. Example 2.2 (Two-sided mean). A natural generalization of Example 2.1 is to consider bivariate random variables. To this end, let (X i,y i ), i = 1,...,n, be an i.i.d. sequence of random variables with distribution P on R 2. Let μ X (P ) denote the mean of the first component of the distribution P and μ Y (P ) the mean of the second component of the distribution P. The parameter of interest, θ 0, is known to satisfy μ X (P ) θ 0 μ Y (P ). For example, θ 0 might be the mean of another distribution Q on R that is known to satisfy μ X (P ) μ(q) μ Y (P ). The identified set is therefore given by Θ 0 (P ) ={θ R : μ X (P ) θ μ Y (P )}. This set may be characterized as the set of minimizers of Q(θ, P ) = (μ X (P ) θ) (θ μ Y (P )) 2 +. The sample analog of Q(θ, P ) is given by ˆQ n (θ) = ( X n θ) (θ Ȳ n ) 2 +. Remark 2.1. Let W i [0, 1] and D i {0, 1}. For example, W i may be the answer to the question, Do you vote Republican or Democrat?, and D i may be the indicator variable for whether the person asked chooses to answer the question. Suppose the researcher observes an i.i.d. sequence (W i D i,d i ), i = 1,...,n, with distribution P ; i.e., W i is observed if and only if D i = 1. The parameter of interest, θ 0 = E{W i }, is not determined by the distribution of the observed data, but the researcher can say with certainty that θ 0 satisfies E P {W i D i } θ 0 E P {W i D i + 1 D i }. By identifying X i = W i D i and Y i = W i D i + 1 D i, this example can be seen to be a special case of Example 2.2. This example is analyzed in detail by Imbens and Manski (2004). See Stoye (2007) for further analysis. Example 2.3 (Regression with interval outcomes). The following example allows for inference in a linear regression model in which the dependent variable is interval-censored. Let (X i,y 1,i,Y 2,i,Yi ), i = 1,...,n, be an i.i.d. sequence of random variables with distribution Q on R k R R R. The parameter of interest, θ 0, is known to satisfy E Q {Yi X i}=x i θ 0, but Yi is unobserved, which precludes conventional estimation of θ 0. Let P denote the distribution of the observed random variables (X i,y 1,i,Y 2,i ). The random variables (Y 1,i,Y 2,i ) are known to satisfy Y 1,i Yi Y 2,i with probability 1 under Q. Thus, θ 0 Θ 0 (P ) ={θ R k : E P {Y 1,i X i } X i θ E P {Y 2,i X i }P -a.s.}. This set may be characterized as the set of minimizers of Q(θ, P ) = E P {(E P {Y 1,i X i } X i θ)2 + + (X i θ E P {Y 2,i X i }) 2 + }. The sample analog ˆQ n (θ) of Q(θ, P ) is given by replacing expectations with appropriately defined estimators, the nature of which may depend on further assumptions about P. Manski and Tamer (2002) characterize the identified set in this setting and also consider the case where Yi is observed, but X i is interval-censored.

5 2790 J.P. Romano, A.M. Shaikh / Journal of Statistical Planning and Inference 138 (2008) Remark 2.2. A Tobit-like model is a special case of the previous example if we suppose further that Y 2,i = Yi and Y 1,i = Yi if Yi > 0, and Y 2,i = 0 and Y 1,i = (or some large negative number if there is a plausible lower bound on Yi ) if Y i 0. It is worthwhile to note that conventional approaches to inference for such a model, while enforcing identification, rely on the much stronger assumption that ε i = Yi X i θ 0 is independent of X i. By allowing for only partial identification, we will be able to draw inferences under much weaker assumptions. Example 2.4 (Moment inequalities). Consider the following generalization of Examples 2.1 and 2.2. Let X i,i = 1,...,n, be an i.i.d. sequence of random variables with distribution P on R k. For j = 1,...,m, let g j (x, θ) be a real-valued function on R k R l. The identified set is assumed to be Θ 0 (P ) ={θ R l : E P {g j (X i, θ)} 0 j s.t. 1 j m}. This set may be characterized as the set of minimizers of Q(θ, P ) = (E P {g j (X i, θ)}) j m or equivalently as {θ Θ : Q(θ, P ) = 0}. The sample analog of Q(θ, P ) is given by ˆQ n (θ) = 1 j m 1 n 1 i n 2 g j (X i, θ). + This choice of Q(θ, P ) is especially noteworthy because it is used by Cilberto and Tamer (2004), Benkard et al. (2007) and Borzekowski and Cohen (2005) in their empirical applications. Since any equality restriction may be thought of as two inequality restrictions, this example may also be viewed as a generalization of the method of moments to allow for equality and inequality restrictions on the moments, rather than just equality restrictions. Remark 2.3. A prominent example of an econometric model which gives rise to moment inequalities is an entry model. See, for example, Andrews et al. (2004) or Cilberto and Tamer (2004) for a detailed description of such models and a derivation of the inequalities. Briefly, consider an entry model with two firms and let X i be the indicator for the event firm 1 enters. Because of the multiplicity of Nash equilibria, the model only gives upper and lower bounds, L(θ) and U(θ), on the probability of this event as a function of the unknown parameter, θ, of the econometric model. It is therefore natural to use the functions g 1 (X i, θ) = L(θ) X i, g 2 (X i, θ) = X i U(θ) as a basis for inference in such a model. 3. Confidence regions for identifiable parameters In this section, we consider the problem of constructing confidence regions for identifiable parameters. We begin in Section 3.1 by recalling a general result about subsampling that will be used frequently throughout the remainder of the paper. The proof of the result, which is lengthy, can be found in Romano and Shaikh (2006a). In Section 3.2, we then treat the construction of sets satisfying (2), before turning our attention in Section 3.3 to the problem of constructing sets satisfying (3). In Section 3.4, we generalize these constructions to confidence regions for functions of identifiable parameters A useful result about subsampling The following theorem describes conditions under which confidence sets constructed using subsampling are asymptotically valid for a fixed P as well as uniformly over P. The conditions for a fixed P are distinct from those described in Politis et al. (1999) and can accommodate certain degenerate situations not covered by the results presented there. As we will see below, this feature of the conditions will be essential for the types of problems we are considering.

6 J.P. Romano, A.M. Shaikh / Journal of Statistical Planning and Inference 138 (2008) Theorem 3.1. Let X i,i= 1,...,n, be an i.i.d. sequence of random variables with distribution P. Let ϑ(p ) be a realvalued parameter and let ˆϑ n = ˆϑ n (X 1,...,X n ) be some estimator of ϑ(p ) that is symmetric in its arguments. Denote by J n (, P ) the distribution of the root τ n ( ˆϑ n ϑ(p )), where τ n is a sequence of known constants. Let b = b n <nbe a sequence of positive integers tending to infinity, but satisfying b/n 0. Let N n = ( n b ) and L n (x) = 1 I{τ b ( ˆϑ n,b,i ϑ(p )) x}, (7) N n 1 i N n where ˆϑ n,b,i denotes the estimate ˆϑ n evaluated at the ith subset of data of size b. Finally, let L 1 n (1 α)= inf{x R : L n(x) 1 α}. Then, for α (0, 1), the following statements are true: (i) If lim sup sup x R {J b (x, P ) J n (x, P )} 0, then lim inf P {τ n( ˆϑ n ϑ(p )) L 1 n (1 α)} 1 α. (ii) If lim sup sup x R {J n (x, P ) J b (x, P )} 0, then lim inf P {τ n( ˆϑ n ϑ(p )) L 1 n (α)} 1 α. (iii) If lim sup x R J b (x, P ) J n (x, P ) =0, then { ( α ( lim inf P L 1 n τ n ( ˆϑ n ϑ(p )) L 2) 1 n 1 α )} 1 α. 2 Moreover, we have that: (iv) If lim sup sup P P sup x R {J b (x, P ) J n (x, P )} 0, then lim inf inf P {τ n( ˆϑ n ϑ(p )) L 1 n (1 α)} 1 α. P P (v) If lim sup sup P P sup x R {J n (x, P ) J b (x, P )} 0, then lim inf inf P {τ n( ˆϑ n ϑ(p )) L 1 n (α)} 1 α. P P (vi) If lim sup P P sup x R J b (x, P ) J n (x, P ) =0, then { ( α ( lim inf inf P L 1 n τ n ( ˆϑ n ϑ(p )) L P P 2) 1 n 1 α )} 1 α. 2 Remark 3.1. The subsampling distribution L n (x) defined by (7) differs from the typical definition given in Politis et al. (1999) in that it centers about the true value of the parameter ϑ(p ) instead of its estimate ˆϑ n. Strictly speaking, it is therefore not possible to compute L n (x) as it is defined above. We may, however, compute L n (x) under the null hypothesis that ϑ(p ) equals a particular value. For our purposes below, this will suffice Pointwise consistency in level We assume in the remainder of the paper that inf θ Θ Q(θ, P ) = 0. Note that this restriction is without loss of generality, as it can always be enforced simply by subtracting the quantity inf θ Θ Q(θ, P ) from the objective function whenever inf θ Θ Q(θ, P ) exists. Our construction exploits the well-known duality between 1 α confidence regions and tests of the individual null hypotheses H θ : Q(θ, P ) = 0 for θ Θ (8)

7 2792 J.P. Romano, A.M. Shaikh / Journal of Statistical Planning and Inference 138 (2008) that control the usual probability of a Type 1 error at level α. The hypotheses H θ are one-sided in the sense that the alternative hypotheses are understood to be K θ : Q(θ, P ) > 0 for θ Θ. Typically, Θ will be a subset of Euclidean space, in which case there may be an uncountably infinite number of hypotheses in this family. For each null hypothesis H θ, a test statistic is required such that large values of the test statistic provide evidence against the null hypothesis. We will use the statistic a n ˆQ n (θ) for some sequence a n for this purpose. The construction of critical values for these tests will be based on subsampling. In order to define the critical values precisely, some further notation is required. Let b = b n <nbe a sequence of positive integers tending to infinity, but satisfying b/n 0. Let N n = ( n b ) and let ˆQ n,b,i (θ) denote the statistic ˆQ n (θ) evaluated at the ith subset of data of size b from the n observations. For α (0, 1), define ˆd n (θ, 1 α) = inf x : 1 I{a b ˆQ n,b,i (θ) x} 1 α N n. (9) 1 i N n We will test each null hypothesis H θ by comparing the test statistic a n ˆQ n (θ) with the critical value ˆd n (θ, 1 α). The set of θ values corresponding to the accepted hypotheses from this procedure will form C n, i.e., C n ={θ Θ : a n ˆQ n (θ) ˆd n (θ, 1 α)}. (10) We now provide conditions under which C n defined by (10) satisfies the desired coverage property (2). Theorem 3.2. Let X i,i = 1,...,n, be an i.i.d. sequence of random variables with distribution P and let b = b n <n be a sequence of positive integers tending to infinity, but satisfying b/n 0. Denote by J n (, θ, P ) the distribution of a n ( ˆQ n (θ) Q(θ, P )). Suppose that for every θ Θ 0 (P ) lim sup sup{j b (x, θ, P ) J n (x, θ, P )} 0. (11) x R Then, C n defined by (10) satisfies (2). Proof. Fix θ Θ 0 (P ). By Theorem 3.1(i), we have immediately that lim inf P {a n ˆQ n (θ) ˆd n (θ, 1 α)} 1 α. Thus, by the (10), it follows that (2) holds. Remark 3.2. It may be useful to consider confidence regions based on critical values given by d n (θ, 1 α) = inf x : 1 I{a b ( ˆQ n,b,i (θ) ˆQ n (θ)) x} 1 α N n 1 i N n (12) instead of (9). Such confidence regions will satisfy the coverage requirement (2) under the assumptions of Theorem 3.2 and the additional assumption that a b /a n 0. The critical values (12) will typically be smaller than the ones given by (9), so the resulting confidence regions will typically be smaller as well. Remark 3.3. Because ( n b ) may be large, it is often more practical to use the following approximation to (9). Let B n be a positive sequence of numbers tending to as n and let I 1,...,I Bn be chosen randomly with or without replacement from the numbers 1,...,N n. Then, it follows from Corollary of Politis et al. (1999) that one may

8 J.P. Romano, A.M. Shaikh / Journal of Statistical Planning and Inference 138 (2008) approximate (9) by inf x : 1 I{a b ˆQ n,b,ii (θ) x} 1 α B n 1 i B n without affecting the conclusions of Theorem 3.2. We now revisit each of the examples described in Section 2 and provide conditions under which the hypothesis (11) of Theorem 3.2 holds. It follows that for each of these examples C n defined by (10) satisfies (2). Example 3.1 (One-sided mean). Recall the setup of Example 2.1. Let a n = n and suppose P is such that the variance σ 2 (P ) exists. We will verify that the required condition (11) holds for any such P. Consider θ Θ 0 (P ). We may assume without loss of generality that μ(p )=0. First note that because ˆQ n (θ) 0, we may restrict attention to x 0 in (11). There are two cases to consider: θ > 0 and θ=0. In the former case, we have for any ε > 0 that X n θ < ε θ with probability approaching one. By choosing ε sufficiently small, we therefore have that a n ( ˆQ n (θ) Q(θ, P )) = n( X n θ) 2 + = 0 with probability approaching 1. Hence, (11) holds trivially. In the latter case, a n ( ˆQ n (θ) Q(θ, P )) = n( X n ) 2 +. Thus, for x 0, J n (x, θ, P ) = P {n( X n ) 2 + x}=p { n X n x}. Since n X n L Φσ(P ), where Φ σ(p ) is the distribution of a mean-zero normal random variable with variance σ 2 (P ), Polya s Theorem implies that sup P { n X n x} Φ σ(p ) (x) 0. x R The desired condition (11) now follows from an appeal to the triangle inequality. Example 3.2 (Two-sided mean). Recall the setup of Example 2.2. Let a n = n and suppose P is such that the variances σ 2 X (P ) and σ2 Y (P ) exist and μ X(P ) μ Y (P ). Denote by σ X,Y (P ) the covariance of X i and Y i under P. We will verify that the required condition (11) holds for any such P. Consider θ Θ 0 (P ). First note, as before, that because ˆQ n (θ) 0, we may restrict attention to x 0 in (11). There are two cases to consider: either μ X (P ) < μ Y (P ) or μ X (P ) = μ Y (P ). If μ X (P ) < μ Y (P ), then we must consider the case in which θ (μ X (P ), μ Y (P )) and the case in which θ {μ X (P ), μ Y (P )} separately. In the former case, it is easy to see that a n ( ˆQ n (θ) Q(θ, P ))=n( X n θ) 2 + +n(θ Ȳ n ) 2 + =0 with probability approaching 1. Hence, (11) holds trivially. In the latter case, we may assume without loss of generality that θ = μ X (P ) and μ X (P ) = 0; the case in which θ = μ Y (P ) is symmetric. In this case, a n ( ˆQ n (θ) Q(θ, P )) = n( X n ) n( Ȳ n ) 2 +. Because μ Y (P ) > 0, we have that with probability tending to 1, a n ( ˆQ n (θ) Q(θ, P )) = n( X n ) 2 +. The desired condition (11) now follows from the reasoning given in Example 3.1. If μ X (P ) = μ Y (P ), then we may assume without loss of generality that this common value is 0. Then, a n ( ˆQ n (θ) Q(θ, P )) = n( X n ) n( Ȳ n ) 2 +. Thus, for x 0, J n (x, θ, P ) = P {n( X n ) n( Ȳ n ) 2 + x}=p {( n X n, nȳ n ) S x }, where S x is the appropriate convex set. Note that ( n X n, nȳ n ) L Φ σx (P ),σ Y (P ), σ X,Y (P ), where Φ σx (P ),σ Y (P ), σ X,Y (P ) is a mean-zero bivariate normal distribution with variances σ 2 X (P ) and σ2 Y (P ) and covariance σ X,Y (P ). It follows from Theorem 2.11 of Bhattacharya and Rao (1976) that sup P {( n X n, nȳ n ) S} Φ σx (P ),σ Y (P ), σ X,Y (P )(S) 0, where S is the set of all convex subsets of R 2. Hence, the desired condition (11) again follows from an appeal to the triangle inequality.

9 2794 J.P. Romano, A.M. Shaikh / Journal of Statistical Planning and Inference 138 (2008) Example 3.3 (Regression with interval outcomes). Recall the setup of Example 2.3. Let a n = n and let {x 1,...,x J } be a set of vectors in R k. For l {1, 2} and j {1,...,J}, let τ l (x j, P ) = E P {Y l,i X i = x j }, (13) ˆτ l (x j ) = 1 n(x j ) where n(x j ) = {1 i n : X i = x j }. Let ˆQ n (θ) = 1 j J 1 i n:x i =x j Y l,i, (14) n(x j ) n {(ˆτ 1(x j ) x j θ)2 + + (x j θ ˆτ 2(x j )) 2 + }. (15) Suppose P is such that (i) supp P (X i ) ={x 1,...,x J } and (ii) the variances of Y 1 and Y 2, σ 2 1 (P ) and σ2 2 (P ), exist. Note that it follows from the above assumptions on P that the vector whose components are given by n(x j )(ˆτ l (x j ) τ l (x j, P )) for j = 1,...,J converges in distribution to a multivariate normal distribution. We will now verify that the required condition (11) holds for this setup. Consider θ Θ 0 (P ). First note, as before, that because ˆQ n (θ) 0, we may restrict attention to x 0 in (11). There are two cases to consider: either θ int(θ 0 (P )) or θ Θ 0 (P ). In the former case, τ 1 (x j, P ) < x j θ < τ 2(x j, P ) for all j. Thus, it is easy to see that a n ˆQ n (θ) = 0 with probability tending to 1. Hence, (11) holds trivially. In the latter case, let I l denote the set of j indices such that τ l (x j, P ) = x j θ. Note that at most one of I 1 or I 2 may be empty. Then, with probability approaching 1, we have that a n ˆQ n (θ) = ( n(xj ) 2 )(ˆτ 1 (x j ) x j θ) + ( n(xj ) 2 )(x j θ ˆτ 2(x j )). + + j I 1 j I 2 Let Ŵ n (θ) denote the vector whose components are given by n(x j )(ˆτ 1 (x j ) x j θ) for j I 1 and n(x j )(x j θ ˆτ 2(x j )) for j I 2 and let W(θ) denote the limiting multivariate normal distribution of Ŵ n (θ). For x 0, we have that J n (x, θ, P ) = P {Ŵ n (θ) S x }, where S x is the appropriate set in S, defined by (24). It follows from Theorem 2.11 of Bhattacharya and Rao (1976) that sup P {Ŵ n (θ) S} P {W(θ) S} 0, where S is the set of all convex subsets of R I 1 + I 2. Hence, the desired condition (11) again follows from an appeal to the triangle inequality. Example 3.4 (Moment inequalities). Recall the setup of Example 2.4. Let a n = n and suppose P is such that the variance of the vector whose components are given by g j (X i, θ) for j = 1,...,mexists for all θ Θ. It follows from this assumption that the vector whose components are given by 1 n 1 i n (g j (X i, θ) E P {g j (X i, θ)}) for j = 1,...,m converges in distribution to a multivariate normal random variable for each θ Θ. We will verify that the required condition (11) holds for any such P. Consider θ Θ 0 (P ). There are two cases to consider: either θ int(θ 0 (P )) or θ Θ 0 (P ). In the former case, E P {g j (X i, θ)} < 0 for all j, so a n ˆQ n (θ) = 0 with probability tending to 1. Hence, (11) holds trivially. In the latter case, let I denote the set of j indices such that E P {g j (X i, θ)}=0. Note that I must be nonempty. Then, with probability approaching one, we have that a n ˆQ n (θ) = 2 1 g j (X i, θ). n j I 1 i n +

10 J.P. Romano, A.M. Shaikh / Journal of Statistical Planning and Inference 138 (2008) Let Ŵ n (θ) denote the vector whose components are given by 1 g j (X i, θ) n 1 i n for j I and let W(θ) denote the limiting multivariate normal distribution. For x 0, we have that J n (x, θ, P ) = P {Ŵ n (θ) S x }, where S x is the appropriate convex set. It follows from Theorem 2.11 of Bhattacharya and Rao (1976) that sup P {Ŵ n (θ) S} P {W(θ) S} 0, where S is the set of all convex subsets of R I. Hence, the desired condition (11) again follows from an appeal to the triangle inequality. Remark 3.4. In each of the examples from Section 2, there are many other choices of Q(θ, P ) sharing the feature that Θ 0 (P ) = arg min Q(θ, P ). (16) θ Θ Consider Example 2.4. Let g(x, θ) denote the vector whose components are given by g j (x, θ), j =1,...,mand extend the definition of ( ) + on the real line to vectors in the natural way. It is easy to see that Q(θ, P ) = (E P {g(x, θ)}) + W(θ)(E P {g(x, θ)) + }, where W(θ) is a positive definite matrix for each θ Θ also satisfies (16) for this example. The objective function given by Q(θ, P ) = max w j (θ)(e P {g j (X, θ)}) +, 1 j m where w j (θ), j = 1,...,m are positive for each θ Θ, also satisfies (16) for this example. Each of these choices of Q(θ, P ) will lead to different test statistics a n ˆQ n (θ), which in turn will lead to different tests of the individual null hypotheses in (8). Of course, more powerful tests will reject more false hypotheses and thereby lead to smaller confidence regions, so the choice of Q(θ, P ) is an important direction for future research. Remark 3.5. If it was the case that under the null hypothesis H θ the statistic a n ˆQ n (θ) converged in distribution to a random variable that was continuous at its 1 α quantile, then it would follow immediately from Theorem of Politis et al. (1999) that C n defined by (10) satisfies (2). Unfortunately, in the examples from Section 2, this assertion does not always hold. Indeed, in many instances, for θ int(θ 0 (P )), it is true that a n ˆQ n (θ) = 0 with probability tending to 1. For this reason, it is necessary to appeal to Theorem 3.1 to establish the desired conclusions. Remark 3.6. Chernozhukov et al. (2007) also consider the problem of constructing confidence regions C n satisfying (2). In an earlier version of their paper (Chernozhukov et al., 2004), they propose a construction that relies upon an initial estimate of Θ 0 (P ). Their estimate of Θ 0 (P ) is given by ˆΘ 0,n ={θ Θ : ˆQ n (θ)<ε n }, where ε n is a positive sequence of constants tending to zero sufficiently slowly. Because of this rate restriction, they are able to show that ˆΘ 0,n satisfies P {Θ 0 (P ) ˆΘ 0,n } 1. They then propose the confidence region C n = θ Θ : a n ˆQ n (θ) sup ˆd n (θ, 1 α) θ ˆΘ, 0,n

11 2796 J.P. Romano, A.M. Shaikh / Journal of Statistical Planning and Inference 138 (2008) where ˆd n (θ, 1 α)} is given by (9). Of course, such a construction is clearly more conservative than the one given by (10). Moreover, the level of ε n is completely arbitrary and the confidence region resulting from application of their method may thus be very large or very small depending on the choice of ε n. For this reason, after learning of the results in an earlier version of our paper, Chernozhukov et al. (2007) suggested the confidence region C n ={θ Θ : a n ˆQ n (θ) min sup ˆd n (θ, 1 α), ˆd n (θ, 1 α) θ ˆΘ, 0,n where ˆd n (θ, 1 α)} is again given by (9). In a first-order asymptotic sense, this confidence region is equivalent to our confidence region (10). Moreover, the extent to which this construction is smaller than ours in finite-samples depends again on the level of ε n, which is left unspecified. For this reason, we prefer our original construction, which does not suffer from such arbitrariness. Remark 3.7. In certain applications, it may be of interest to test the null hypothesis that there exists some point θ Θ for which Q(θ, P ) = 0. Such a test may be viewed as a specification test of the model. One may test such a hypothesis by first constructing a confidence region for identifiable parameters and then rejecting the hypothesis if the confidence region is empty. To see this, note that under the null hypothesis, Θ 0 (P ) =. Thus, there is some θ Θ 0 (P ). Under the assumptions of Theorem 3.2, it must be the case that lim inf P {θ C n } 1 α. Thus, under the null hypothesis C n will be nonempty with probability at least 1 α asymptotically. Remark 3.8. Our construction of critical values for testing H θ : Q(θ, P ) = 0 versus K θ : Q(θ, P ) > 0 has used subsampling. We now show by example that the bootstrap may fail to approximate the distribution of the test statistic, a n ˆQ n (θ), under the null hypothesis. To this end, recall the setup of Example 2.1. Let a n =n and suppose that P =N(0, 1). Note that Q(θ, P )=0 at θ=0, so the null hypothesis is true when θ=0. Our argument follows the one given by Andrews (2000) to show that the bootstrap is inconsistent when the parameter is on the boundary of the parameter space. Note that a n ˆQ n (0) = ( n X n ) 2 + (Z)2 +, where Z is a standard normal random variable. Denote by Xi, i = 1,...,nan i.i.d. sequence of random variables with distribution Pˆ n given by the empirical distribution of the original observations X i, i = 1,...,n. Define ˆQ n (θ) = ( X n θ)2 + and consider the event { A c = ω : lim inf n X n < c} for c (0, ). The Law of the Iterated Logarithm asserts that (17) lim sup n X n 2 log log n = 1 P a.s. This in turn implies that { n X n > 2 log log n i.o.} and by symmetry that { n X n < 2 log log n i.o.} It follows that P {A c }=1. P a.s. P a.s.

12 J.P. Romano, A.M. Shaikh / Journal of Statistical Planning and Inference 138 (2008) Now consider the bootstrap approximation to the distribution of a n ˆQ n (θ) at θ = 0, which is given by L(a n ( ˆQ n (0) ˆQ n (0)) X i,i= 1,..., n). (18) This approximation mimics the hypothesis that Q(0, P ) = 0 by centering ˆQ n (0) about ˆQ n (0). For ω A c, consider a subsequence n k of n 1 for which n k X nk (ω)< c for all k. For such a subsequence, we have, conditionally on X i (ω), i = 1,...,n, that n k ( ˆQ n k (0) ˆQ nk (0)(ω)) = ( n k X n k ) 2 + = ( n k ( X n k X nk (ω)) + n k X nk (ω)) 2 + ( n k ( X n k X nk (ω)) c) 2 + L (Z c) 2 +, where Z is again a standard normal random variable. The first equality and the inequality follow from the fact that nk X nk (ω)< c<0. It follows that the bootstrap fails to approximate the distribution of a n ˆQ n (θ) at θ = 0. It is worthwhile to consider the actual probability of a Type 1 error if the bootstrap approximation above is used instead of subsampling. To this end, first note that the probability of such an error is given by P {a n ˆQ n (0)>ĉ boot,n (1 α)}, (19) where ĉ boot,n (1 α) is the 1 α quantile of (18). Recall that a n ˆQ n (0) = ( n X n ) 2 + (Z)2 +, where Z is a standard normal random variable. Note that a n ( ˆQ n (0) ˆQ n (0)) = ( n( X n X n ) + n X n ) 2 + ( n X n ) 2 +. Therefore, (18) converges in distribution to L((Z + Z) 2 + (Z)2 + Z), (20) where Z is a standard normal random variable distributed independently of Z. The probability (19) is therefore asymptotically equal to P {(Z) 2 + > c(1 α Z)}, where c(1 α Z) is the 1 α quantile of (20). Using this representation, we simulate the asymptotic value of (19) and find that the probability of Type 1 error is controlled, but too conservatively for practical purposes. In fact, with 2000 simulations, we estimate that (19) is asymptotically equal to 0 to three significant digits for both α =.1 and α = Uniform consistency in level We now turn our attention to providing conditions under which C n defined by (10) satisfies (3). Recall from the discussion in the introduction that inferences made on the basis of asymptotic arguments with confidence regions that are not uniformly consistent in level may be very misleading in the sense that asymptotic approximations may be poor for arbitrarily large sample sizes. Theorem 3.3. Let X i,i = 1,...,n, be an i.i.d. sequence of random variables with distribution P and let b = b n <n be a sequence of positive integers tending to infinity, but satisfying b/n 0. Denote by J n (, θ, P ) the distribution of a n ( ˆQ n (θ) Q(θ, P )). Suppose that lim sup sup θ Θ sup P P:θ Θ 0 (P ) x R Then, C n defined by (10) satisfies (3). sup{j b (x, θ, P ) J n (x, θ, P )} 0. (21)

13 2798 J.P. Romano, A.M. Shaikh / Journal of Statistical Planning and Inference 138 (2008) Proof. By Theorem 3.1(iv), we have immediately that lim inf inf inf P {a n ˆQ n (θ) ˆd n (θ, 1 α)} 1 α. θ Θ P P:θ Θ 0 (P ) Thus, by definition (10), it follows that (3) holds. We now apply Theorem 3.3 to construct confidence regions satisfying the coverage requirement (3) for each of the examples described in Section 2. Example 3.5 (One-sided mean). Recall the setup of Example 2.1. We will now use Theorem 3.3 to show that for this example C n defined by (10) satisfies (3) for a large class of distributions P. To this end, let a n = n and let P be a set of distributions satisfying lim sup λ P P [ X μ(p ) 2 E P σ 2 I (P ) { X μ(p ) σ(p ) }] > λ = 0. (22) We will argue by contradiction that the required condition (21) holds. If the result were false, then there would exist a subsequence n j and a corresponding sequence (θ nj,p nj ) Θ P such that μ(p nj ) θ nj and sup{j bnj (x, θ nj,p nj ) J nj (x, θ nj,p nj )} δ (23) x R for some δ > 0. Since a n ˆQ n (θ nj ) 0, we may restrict attention to x 0. We may thus rewrite (23) as where sup{p nj {(Z bnj (P nj ) + T bnj (θ nj,p nj )) 2 + x} P n j {(Z nj (P nj ) + T nj (θ nj,p nj )) 2 + x}}, x 0 Z n (P ) = n( X n μ(p )), T n (θ, P ) = n(μ(p ) θ). Since b nj n j and μ(p nj ) θ nj, we have that T bnj (θ nj,p nj ) T nj (θ nj,p nj ). Thus, (23) is bounded above by sup{p nj {(Z bnj (P nj ) + T nj (θ nj,p nj )) 2 + x} P n j {(Z nj (P nj ) + T nj (θ nj,p nj )) 2 + x}}. x 0 By redefining x appropriately and absorbing the term T nj (θ nj,p nj ) into x, it follows that: sup{p nj {Z bnj (P nj ) x} P nj {Z nj (P nj ) x}} 0. x R By Lemma of Lehmann and Romano (2005), this yields the desired contradiction. Example 3.6 (Two-sided mean). Recall the setup of Example 2.2. As in Example 3.5, we may use Theorem 3.3 to show that for this example C n defined by (10) satisfies (3) for a large class of distributions P. To this end, let a n = n, let P be a set of bivariate distributions such that the marginal distributions satisfy (22) and μ X (P ) μ Y (P ). As before, we will argue by contradiction that the required condition (21) holds, but we will require the following result, which generalizes Lemma of Lehmann and Romano (2005), in order to do so. For the sake of continuity, the proof of the result is found in Appendix A. Lemma 3.1. Let P be a set of distributions on R k such that the marginal distributions satisfy (22). Let X i = (X 1,i,...,X k,i ), i = 1,...,n be an i.i.d. sequence of random variables with distribution P P. Denote by Φ V ( ) the probability distribution of a multivariate normal random variable with mean 0 and variance V and by Σ(P ) the variance of P. Then, sup sup P {( n( X 1,n μ 1 (P ),..., n( X k,n μ k (P )) S} Φ Σ(P ) (S) 0, P P

14 J.P. Romano, A.M. Shaikh / Journal of Statistical Planning and Inference 138 (2008) where S ={S R k : S convex and Φ V ( S) = 0 for all p.s.d. V s.t. V i,i = 1 for 1 i k}. (24) Proof. See Appendix A. We now return to verifying (21). If the result were false, then there would exist a subsequence n j and a corresponding sequence (θ nj,p nj ) Θ P such that μ X (P nj ) θ nj μ Y (P nj ) and sup{j bnj (x, θ nj,p nj ) J nj (x, θ nj,p nj )} δ (25) x R for some δ > 0. Note that a bnj ˆQ bnj (θ nj ) where (Z 1,bnj (P nj ) + T 1,bnj (θ nj,p nj )) (Z 2,b nj (P nj ) + T 2,bnj (θ nj,p nj )) 2 +, Z 1,n (P ) = n( X n μ X (P )), Z 2,n (P ) = n(μ Y (P ) Ȳ n ), T 1,n (θ, P ) = n(μ X (P ) θ), T 2,n (θ, P ) = n(θ μ Y (P )). Let J bnj (x, θ nj,p nj ) be the distribution of the statistic (Z 1,bnj (P nj ) + T 1,nj (θ nj,p nj )) (Z 2,b nj (P nj ) + T 2,nj (θ nj,p nj )) 2 +. Since T k,bnj (θ nj,p nj ) T k,nj (θ nj,p nj ), we have that J bnj (x, θ nj,p nj ) J bnj (x, θ nj,p nj ). Thus, (25) implies that sup{ J bnj (x, θ nj,p nj ) J nj (x, θ nj,p nj )} 0. (26) x R Since a n ˆQ n (θ) 0, we may restrict attention to x 0. For such x, J bnj (x, θ nj,p nj ) is simply the probability under P nj that the vector (Z 1,bnj (P nj ), Z 2,bnj (P nj )) lies in a set S x S, where S is defined by (24). Importantly, J nj (x, θ nj,p nj ) is simply the probability under P nj that the vector (Z 1,nj (P nj ), Z 2,nj (P nj )) lies in the same set. But, by Lemma 3.1, however, we know that and sup P nj {(Z 1,nj (P nj ), Z 2,nj (P nj )) S} Φ σx (P nj ),σ Y (P nj ),ρ(p nj )(S) 0 sup P nj {(Z 1,bnj (P nj ), Z 2,bnj (P nj )) S} Φ σx (P nj ),σ Y (P nj ),ρ(p nj )(S) 0. An appeal to the triangle inequality yields the desired contradiction to (26). Remark 3.9. Recall that Imbens and Manski (2004) analyze a special case of Example 2.2 in which X i = W i D i and Y i = W i D i + 1 D i where W i [0, 1] and D i {0, 1}. The motivation for their study of this problem stems from considering [ C n = X n z 1 α ˆσ X,n, Ȳ n + z ] 1 α ˆσ Y,n, (27) n n

15 2800 J.P. Romano, A.M. Shaikh / Journal of Statistical Planning and Inference 138 (2008) where ˆσ X,n and ˆσ Y,n are the usual consistent estimators of σ X (P ) and σ Y (P ), respectively, and z 1 α is the 1 α quantile of the standard normal distribution. For any P such that P {D i = 1} < 1, C n defined by (27) satisfies (2). To see this, consider the two cases θ int(θ 0 (P )) and θ Θ 0 (P ) separately. In the former case, X n < θ < Ȳ n with probability approaching 1, which implies that θ C n with probability approaching 1 as well. In the latter case, θ = μ X (P ) or θ = μ Y (P ). Suppose θ = μ X (P ); the case in which θ = μ Y (P ) is completely symmetric. Then, θ < Ȳ n with probability approaching 1 and θ > X n z 1 α ˆσ X,n / n with probability at least 1 α asymptotically. We therefore have that θ C n with probability at least 1 α asymptotically, as desired. Imbens and Manski (2004) show, however, that this convergence is not uniform over all P satisfying P {D i = 1} < 1. Specifically, they show that for every sample size n there is a P with P {D i = 1} sufficiently close, but not equal, to 1 and a θ Θ 0 (P ) for which P {θ C n } 1 2α. To rectify this shortcoming, Imbens and Manski (2004) propose an alternative to C n defined by (27) which satisfies the uniform coverage requirement (3) for P = P such that inf P P P {D i = 1} > 0, inf P P σ2 W i D i =1 (P ) > 0. We may apply the analysis of Example 3.6 and conclude immediately that C n defined by (10) also satisfies (3) for P such that the distributions of X i and Y i defined above satisfy (22). This class of distributions is larger than the one considered by Imbens and Manski (2004). To see this, first note that inf P P σ2 W i D i (P ) inf P P P {D i = 1}σ 2 W i D i =1 (P ) inf P {D P P i = 1} inf P P σ2 W i D i =1 (P ) > 0. Since W i D i is supported on a compact set, it follows that for any δ > 0 { ( Wi ) } D i μ Wi D sup E i (P ) 2+δ P <. P P σ Wi D i (P ) From the pointwise inequality x 2 I{x>λ} x 2+δ λ δ for λ > 0 and δ > 0, we therefore have that (22) holds for W i D i and P = P. Next note that to show that inf P P σ2 W i D i +1 D i (P ) > 0 it suffices to show that inf P P ρ W i D i,1 D i (P ) > 1. (28) But, by direct calculation, we have that 1 ρ Wi D i,1 D i (P ) =. σ 2 W 1 + i D i =1 (P ) E P {W i D i = 1} 2 P {D i = 0} It now follows immediately that (28) holds. We may therefore argue as above for W i D i to establish that (22) holds for W i D i + 1 D i and P = P. Hence, P is a subset of the set of distributions obtained by applying the analysis of Example 3.6.

16 J.P. Romano, A.M. Shaikh / Journal of Statistical Planning and Inference 138 (2008) It is worthwhile to note that the shortcoming of the construction given by (27) pointed out by Imbens and Manski (2004) disappears if we require further that the confidence region C n =[ˆL n, Û n ] is equitailed in the sense that for any θ Θ 0 (P ) lim sup P {θ < ˆL n } α 2, lim sup P {θ > Û n } α 2. One such confidence region is given by (27) in which z 1 α is replaced by z 1 α 2. Note that even if P {D i = 1} =1, so Θ 0 (P ) is a singleton, the confidence region C n defined in this way satisfies (2). Example 3.7 (Regression with interval outcomes). Recall the setup of Example 2.3. Let a n =n and let {x 1,...,x J } be a set of vectors in R k. For l {1, 2} and j {1,...,J}, let τ l (x j, P ) and ˆτ l (x j ) be given by (13) and (14), respectively. Let ˆQ n (θ) be given by (15). Suppose P is a set of distributions such that (i) supp P (X i ) ={x 1,...,x J } for all P P, (ii) forl {1, 2} and j {1,...,J} (Y l,i E P {Y l,i X i = x j })I{X i = x j } (29) satisfies (22), and (iii) inf P P P {X i = x j } > 0. (30) We now argue that the required condition (21) holds under these assumptions. Note that we may rewrite a n ˆQ n (θ) as { ( n(xj ) 2 ( n(xj ) } 2 )(ˆτ 1 (x j ) x j θ) + )(x j θ ˆτ 2(x j )) j m The analysis for this example will therefore be identical to the analysis for Example 3.8 presented below provided that we can show that a central limit theorem holds for the vector whose components are given by n(x j )(ˆτ l (x j ) x j θ) uniformly in P P. To see that this holds under the above assumptions, write n 1 n(xj )(ˆτ l (x j ) τ l (x j, P )) = n(x j ) n 1 i n (Y l,i E P {Y l,i X i = x j })I{X i = x j }. From Lemma 3.1, we have immediately that such a central limit theorem holds for the vector whose components are given by (29). From Chebychev s inequality and the fact that I{X i = x j } is bounded, we also have that n(x j )/n converges in probability to P {X i =x j } uniformly in P P. It follows from (30) that n/n(x j ) converges in probability to 1/P {X i = x j } uniformly in P P. The desired conclusion now follows from Slutsky s theorem. Example 3.8 (Moment inequalities). Recall the setup of Example 2.4. Let a n = n. Denote by Σ(θ, P ) the variance of (g 1 (X i, θ),..., g m (X i, θ)) under P and by σ 2 j (θ, P ) the jth diagonal element of Σ(θ, P ). Suppose P is a set of distributions such that for each 1 j m lim sup sup λ 0 P P θ Θ 0 (P ) [ (gj (X i, θ) E P {g j (X i, θ)}) 2 { gj (X i, θ) E P {g j (X i, θ)} E P σ 2 j (θ, P ) I > λ} ] = 0. σ j (θ, P ) We now argue by contradiction that the required condition (21) holds under this assumption. If the result were false, then there would exist a subsequence n k and a corresponding sequence (θ nk,p nk ) Θ P such that θ nk Θ 0 (P nk ) and sup{j bnk (x, θ nk,p nk ) J nk (x, θ nk,p nk )} δ (31) x R

Inference for Identifiable Parameters in Partially Identified Econometric Models

Inference for Identifiable Parameters in Partially Identified Econometric Models Inference for Identifiable Parameters in Partially Identified Econometric Models Joseph P. Romano Department of Statistics Stanford University romano@stat.stanford.edu Azeem M. Shaikh Department of Economics

More information

On the Uniform Asymptotic Validity of Subsampling and the Bootstrap

On the Uniform Asymptotic Validity of Subsampling and the Bootstrap On the Uniform Asymptotic Validity of Subsampling and the Bootstrap Joseph P. Romano Departments of Economics and Statistics Stanford University romano@stanford.edu Azeem M. Shaikh Department of Economics

More information

ON THE UNIFORM ASYMPTOTIC VALIDITY OF SUBSAMPLING AND THE BOOTSTRAP. Joseph P. Romano Azeem M. Shaikh

ON THE UNIFORM ASYMPTOTIC VALIDITY OF SUBSAMPLING AND THE BOOTSTRAP. Joseph P. Romano Azeem M. Shaikh ON THE UNIFORM ASYMPTOTIC VALIDITY OF SUBSAMPLING AND THE BOOTSTRAP By Joseph P. Romano Azeem M. Shaikh Technical Report No. 2010-03 April 2010 Department of Statistics STANFORD UNIVERSITY Stanford, California

More information

large number of i.i.d. observations from P. For concreteness, suppose

large number of i.i.d. observations from P. For concreteness, suppose 1 Subsampling Suppose X i, i = 1,..., n is an i.i.d. sequence of random variables with distribution P. Let θ(p ) be some real-valued parameter of interest, and let ˆθ n = ˆθ n (X 1,..., X n ) be some estimate

More information

Comparison of inferential methods in partially identified models in terms of error in coverage probability

Comparison of inferential methods in partially identified models in terms of error in coverage probability Comparison of inferential methods in partially identified models in terms of error in coverage probability Federico A. Bugni Department of Economics Duke University federico.bugni@duke.edu. September 22,

More information

Partial Identification and Confidence Intervals

Partial Identification and Confidence Intervals Partial Identification and Confidence Intervals Jinyong Hahn Department of Economics, UCLA Geert Ridder Department of Economics, USC September 17, 009 Abstract We consider statistical inference on a single

More information

Lecture 8 Inequality Testing and Moment Inequality Models

Lecture 8 Inequality Testing and Moment Inequality Models Lecture 8 Inequality Testing and Moment Inequality Models Inequality Testing In the previous lecture, we discussed how to test the nonlinear hypothesis H 0 : h(θ 0 ) 0 when the sample information comes

More information

Practical and Theoretical Advances in Inference for Partially Identified Models

Practical and Theoretical Advances in Inference for Partially Identified Models Practical and Theoretical Advances in Inference for Partially Identified Models Ivan A. Canay Department of Economics Northwestern University iacanay@northwestern.edu Azeem M. Shaikh Department of Economics

More information

Inference for Subsets of Parameters in Partially Identified Models

Inference for Subsets of Parameters in Partially Identified Models Inference for Subsets of Parameters in Partially Identified Models Kyoo il Kim University of Minnesota June 2009, updated November 2010 Preliminary and Incomplete Abstract We propose a confidence set for

More information

A Course in Applied Econometrics. Lecture 10. Partial Identification. Outline. 1. Introduction. 2. Example I: Missing Data

A Course in Applied Econometrics. Lecture 10. Partial Identification. Outline. 1. Introduction. 2. Example I: Missing Data Outline A Course in Applied Econometrics Lecture 10 1. Introduction 2. Example I: Missing Data Partial Identification 3. Example II: Returns to Schooling 4. Example III: Initial Conditions Problems in

More information

SIMILAR-ON-THE-BOUNDARY TESTS FOR MOMENT INEQUALITIES EXIST, BUT HAVE POOR POWER. Donald W. K. Andrews. August 2011

SIMILAR-ON-THE-BOUNDARY TESTS FOR MOMENT INEQUALITIES EXIST, BUT HAVE POOR POWER. Donald W. K. Andrews. August 2011 SIMILAR-ON-THE-BOUNDARY TESTS FOR MOMENT INEQUALITIES EXIST, BUT HAVE POOR POWER By Donald W. K. Andrews August 2011 COWLES FOUNDATION DISCUSSION PAPER NO. 1815 COWLES FOUNDATION FOR RESEARCH IN ECONOMICS

More information

Semi and Nonparametric Models in Econometrics

Semi and Nonparametric Models in Econometrics Semi and Nonparametric Models in Econometrics Part 4: partial identification Xavier d Haultfoeuille CREST-INSEE Outline Introduction First examples: missing data Second example: incomplete models Inference

More information

VALIDITY OF SUBSAMPLING AND PLUG-IN ASYMPTOTIC INFERENCE FOR PARAMETERS DEFINED BY MOMENT INEQUALITIES

VALIDITY OF SUBSAMPLING AND PLUG-IN ASYMPTOTIC INFERENCE FOR PARAMETERS DEFINED BY MOMENT INEQUALITIES Econometric Theory, 2009, Page 1 of 41. Printed in the United States of America. doi:10.1017/s0266466608090257 VALIDITY OF SUBSAMPLING AND PLUG-IN ASYMPTOTIC INFERENCE FOR PARAMETERS DEFINED BY MOMENT

More information

Exact and Approximate Stepdown Methods For Multiple Hypothesis Testing

Exact and Approximate Stepdown Methods For Multiple Hypothesis Testing Exact and Approximate Stepdown Methods For Multiple Hypothesis Testing Joseph P. Romano Department of Statistics Stanford University Michael Wolf Department of Economics and Business Universitat Pompeu

More information

Control of Generalized Error Rates in Multiple Testing

Control of Generalized Error Rates in Multiple Testing Institute for Empirical Research in Economics University of Zurich Working Paper Series ISSN 1424-0459 Working Paper No. 245 Control of Generalized Error Rates in Multiple Testing Joseph P. Romano and

More information

BAYESIAN INFERENCE IN A CLASS OF PARTIALLY IDENTIFIED MODELS

BAYESIAN INFERENCE IN A CLASS OF PARTIALLY IDENTIFIED MODELS BAYESIAN INFERENCE IN A CLASS OF PARTIALLY IDENTIFIED MODELS BRENDAN KLINE AND ELIE TAMER UNIVERSITY OF TEXAS AT AUSTIN AND HARVARD UNIVERSITY Abstract. This paper develops a Bayesian approach to inference

More information

University of California San Diego and Stanford University and

University of California San Diego and Stanford University and First International Workshop on Functional and Operatorial Statistics. Toulouse, June 19-21, 2008 K-sample Subsampling Dimitris N. olitis andjoseph.romano University of California San Diego and Stanford

More information

SIMILAR-ON-THE-BOUNDARY TESTS FOR MOMENT INEQUALITIES EXIST, BUT HAVE POOR POWER. Donald W. K. Andrews. August 2011 Revised March 2012

SIMILAR-ON-THE-BOUNDARY TESTS FOR MOMENT INEQUALITIES EXIST, BUT HAVE POOR POWER. Donald W. K. Andrews. August 2011 Revised March 2012 SIMILAR-ON-THE-BOUNDARY TESTS FOR MOMENT INEQUALITIES EXIST, BUT HAVE POOR POWER By Donald W. K. Andrews August 2011 Revised March 2012 COWLES FOUNDATION DISCUSSION PAPER NO. 1815R COWLES FOUNDATION FOR

More information

The Numerical Delta Method and Bootstrap

The Numerical Delta Method and Bootstrap The Numerical Delta Method and Bootstrap Han Hong and Jessie Li Stanford University and UCSC 1 / 41 Motivation Recent developments in econometrics have given empirical researchers access to estimators

More information

Identification and Inference on Regressions with Missing Covariate Data

Identification and Inference on Regressions with Missing Covariate Data Identification and Inference on Regressions with Missing Covariate Data Esteban M. Aucejo Department of Economics London School of Economics e.m.aucejo@lse.ac.uk V. Joseph Hotz Department of Economics

More information

Partial Identification and Inference in Binary Choice and Duration Panel Data Models

Partial Identification and Inference in Binary Choice and Duration Panel Data Models Partial Identification and Inference in Binary Choice and Duration Panel Data Models JASON R. BLEVINS The Ohio State University July 20, 2010 Abstract. Many semiparametric fixed effects panel data models,

More information

EXACT AND ASYMPTOTICALLY ROBUST PERMUTATION TESTS. Eun Yi Chung Joseph P. Romano

EXACT AND ASYMPTOTICALLY ROBUST PERMUTATION TESTS. Eun Yi Chung Joseph P. Romano EXACT AND ASYMPTOTICALLY ROBUST PERMUTATION TESTS By Eun Yi Chung Joseph P. Romano Technical Report No. 20-05 May 20 Department of Statistics STANFORD UNIVERSITY Stanford, California 94305-4065 EXACT AND

More information

Endogenous binary choice models with median restrictions: A comment

Endogenous binary choice models with median restrictions: A comment Available online at www.sciencedirect.com Economics Letters 98 (2008) 23 28 www.elsevier.com/locate/econbase Endogenous binary choice models with median restrictions: A comment Azeem M. Shaikh a, Edward

More information

and Level Sets First version: August 16, This version: February 15, Abstract

and Level Sets First version: August 16, This version: February 15, Abstract Consistency of Plug-In Estimators of Upper Contour and Level Sets Neşe Yıldız First version: August 16, 2004. This version: February 15, 2008. Abstract This note studies the problem of estimating the set

More information

Fall 2017 STAT 532 Homework Peter Hoff. 1. Let P be a probability measure on a collection of sets A.

Fall 2017 STAT 532 Homework Peter Hoff. 1. Let P be a probability measure on a collection of sets A. 1. Let P be a probability measure on a collection of sets A. (a) For each n N, let H n be a set in A such that H n H n+1. Show that P (H n ) monotonically converges to P ( k=1 H k) as n. (b) For each n

More information

INFERENCE APPROACHES FOR INSTRUMENTAL VARIABLE QUANTILE REGRESSION. 1. Introduction

INFERENCE APPROACHES FOR INSTRUMENTAL VARIABLE QUANTILE REGRESSION. 1. Introduction INFERENCE APPROACHES FOR INSTRUMENTAL VARIABLE QUANTILE REGRESSION VICTOR CHERNOZHUKOV CHRISTIAN HANSEN MICHAEL JANSSON Abstract. We consider asymptotic and finite-sample confidence bounds in instrumental

More information

arxiv: v3 [math.st] 23 May 2016

arxiv: v3 [math.st] 23 May 2016 Inference in partially identified models with many moment arxiv:1604.02309v3 [math.st] 23 May 2016 inequalities using Lasso Federico A. Bugni Mehmet Caner Department of Economics Department of Economics

More information

Lecture 7 Introduction to Statistical Decision Theory

Lecture 7 Introduction to Statistical Decision Theory Lecture 7 Introduction to Statistical Decision Theory I-Hsiang Wang Department of Electrical Engineering National Taiwan University ihwang@ntu.edu.tw December 20, 2016 1 / 55 I-Hsiang Wang IT Lecture 7

More information

The International Journal of Biostatistics

The International Journal of Biostatistics The International Journal of Biostatistics Volume 7, Issue 1 2011 Article 12 Consonance and the Closure Method in Multiple Testing Joseph P. Romano, Stanford University Azeem Shaikh, University of Chicago

More information

Hypothesis Testing. 1 Definitions of test statistics. CB: chapter 8; section 10.3

Hypothesis Testing. 1 Definitions of test statistics. CB: chapter 8; section 10.3 Hypothesis Testing CB: chapter 8; section 0.3 Hypothesis: statement about an unknown population parameter Examples: The average age of males in Sweden is 7. (statement about population mean) The lowest

More information

arxiv: v2 [math.st] 18 Feb 2013

arxiv: v2 [math.st] 18 Feb 2013 The Annals of Statistics 2012, Vol. 40, No. 6, 2798 2822 DOI: 10.1214/12-AOS1051 c Institute of Mathematical Statistics, 2012 arxiv:1204.2762v2 [math.st] 18 Feb 2013 ON THE UNIFORM ASYMPTOTIC VALIDITY

More information

Exact and Approximate Stepdown Methods for Multiple Hypothesis Testing

Exact and Approximate Stepdown Methods for Multiple Hypothesis Testing Exact and Approximate Stepdown Methods for Multiple Hypothesis Testing Joseph P. ROMANO and Michael WOLF Consider the problem of testing k hypotheses simultaneously. In this article we discuss finite-

More information

A Dual Approach to Inference for Partially Identified Econometric Models

A Dual Approach to Inference for Partially Identified Econometric Models A Dual Approach to Inference for Partially Identified Econometric Models Hiroaki Kaido Department of Economics Boston University December, 2015 Abstract This paper considers inference for the set Θ I of

More information

Identification and Inference on Regressions with Missing Covariate Data

Identification and Inference on Regressions with Missing Covariate Data Identification and Inference on Regressions with Missing Covariate Data Esteban M. Aucejo Department of Economics London School of Economics and Political Science e.m.aucejo@lse.ac.uk V. Joseph Hotz Department

More information

INFERENCE FOR PARAMETERS DEFINED BY MOMENT INEQUALITIES USING GENERALIZED MOMENT SELECTION. DONALD W. K. ANDREWS and GUSTAVO SOARES

INFERENCE FOR PARAMETERS DEFINED BY MOMENT INEQUALITIES USING GENERALIZED MOMENT SELECTION. DONALD W. K. ANDREWS and GUSTAVO SOARES INFERENCE FOR PARAMETERS DEFINED BY MOMENT INEQUALITIES USING GENERALIZED MOMENT SELECTION BY DONALD W. K. ANDREWS and GUSTAVO SOARES COWLES FOUNDATION PAPER NO. 1291 COWLES FOUNDATION FOR RESEARCH IN

More information

Introduction to Partial Identification

Introduction to Partial Identification to Partial Identification C. Bontemps 1 1 Toulouse School of Economics DEEQA - Winter 2014 Partial Identification in Econometrics I A first example The traditional approach in statistics and econometrics

More information

Asymptotic Distortions in Locally Misspecified Moment Inequality Models

Asymptotic Distortions in Locally Misspecified Moment Inequality Models Asymptotic Distortions in Locally Misspecified Moment Inequality Models Federico A. Bugni Department of Economics Duke University Ivan A. Canay Department of Economics Northwestern University Patrik Guggenberger

More information

Supplement to Quantile-Based Nonparametric Inference for First-Price Auctions

Supplement to Quantile-Based Nonparametric Inference for First-Price Auctions Supplement to Quantile-Based Nonparametric Inference for First-Price Auctions Vadim Marmer University of British Columbia Artyom Shneyerov CIRANO, CIREQ, and Concordia University August 30, 2010 Abstract

More information

MAT 570 REAL ANALYSIS LECTURE NOTES. Contents. 1. Sets Functions Countability Axiom of choice Equivalence relations 9

MAT 570 REAL ANALYSIS LECTURE NOTES. Contents. 1. Sets Functions Countability Axiom of choice Equivalence relations 9 MAT 570 REAL ANALYSIS LECTURE NOTES PROFESSOR: JOHN QUIGG SEMESTER: FALL 204 Contents. Sets 2 2. Functions 5 3. Countability 7 4. Axiom of choice 8 5. Equivalence relations 9 6. Real numbers 9 7. Extended

More information

Sharp identification regions in models with convex moment predictions

Sharp identification regions in models with convex moment predictions Sharp identification regions in models with convex moment predictions Arie Beresteanu Ilya Molchanov Francesca Molinari The Institute for Fiscal Studies Department of Economics, UCL cemmap working paper

More information

Econometric Analysis of Games 1

Econometric Analysis of Games 1 Econometric Analysis of Games 1 HT 2017 Recap Aim: provide an introduction to incomplete models and partial identification in the context of discrete games 1. Coherence & Completeness 2. Basic Framework

More information

Wooldridge, Introductory Econometrics, 4th ed. Appendix C: Fundamentals of mathematical statistics

Wooldridge, Introductory Econometrics, 4th ed. Appendix C: Fundamentals of mathematical statistics Wooldridge, Introductory Econometrics, 4th ed. Appendix C: Fundamentals of mathematical statistics A short review of the principles of mathematical statistics (or, what you should have learned in EC 151).

More information

Multiple Testing of One-Sided Hypotheses: Combining Bonferroni and the Bootstrap

Multiple Testing of One-Sided Hypotheses: Combining Bonferroni and the Bootstrap University of Zurich Department of Economics Working Paper Series ISSN 1664-7041 (print) ISSN 1664-705X (online) Working Paper No. 254 Multiple Testing of One-Sided Hypotheses: Combining Bonferroni and

More information

Statistical Properties of Numerical Derivatives

Statistical Properties of Numerical Derivatives Statistical Properties of Numerical Derivatives Han Hong, Aprajit Mahajan, and Denis Nekipelov Stanford University and UC Berkeley November 2010 1 / 63 Motivation Introduction Many models have objective

More information

The Delta Method and Applications

The Delta Method and Applications Chapter 5 The Delta Method and Applications 5.1 Local linear approximations Suppose that a particular random sequence converges in distribution to a particular constant. The idea of using a first-order

More information

INVALIDITY OF THE BOOTSTRAP AND THE M OUT OF N BOOTSTRAP FOR INTERVAL ENDPOINTS DEFINED BY MOMENT INEQUALITIES. Donald W. K. Andrews and Sukjin Han

INVALIDITY OF THE BOOTSTRAP AND THE M OUT OF N BOOTSTRAP FOR INTERVAL ENDPOINTS DEFINED BY MOMENT INEQUALITIES. Donald W. K. Andrews and Sukjin Han INVALIDITY OF THE BOOTSTRAP AND THE M OUT OF N BOOTSTRAP FOR INTERVAL ENDPOINTS DEFINED BY MOMENT INEQUALITIES By Donald W. K. Andrews and Sukjin Han July 2008 COWLES FOUNDATION DISCUSSION PAPER NO. 1671

More information

Quantile Regression for Panel Data Models with Fixed Effects and Small T : Identification and Estimation

Quantile Regression for Panel Data Models with Fixed Effects and Small T : Identification and Estimation Quantile Regression for Panel Data Models with Fixed Effects and Small T : Identification and Estimation Maria Ponomareva University of Western Ontario May 8, 2011 Abstract This paper proposes a moments-based

More information

Asymptotically Efficient Estimation of Models Defined by Convex Moment Inequalities

Asymptotically Efficient Estimation of Models Defined by Convex Moment Inequalities Asymptotically Efficient Estimation of Models Defined by Convex Moment Inequalities Hiroaki Kaido Department of Economics Boston University hkaido@bu.edu Andres Santos Department of Economics UC San Diego

More information

Master s Written Examination - Solution

Master s Written Examination - Solution Master s Written Examination - Solution Spring 204 Problem Stat 40 Suppose X and X 2 have the joint pdf f X,X 2 (x, x 2 ) = 2e (x +x 2 ), 0 < x < x 2

More information

Asymptotic Statistics-III. Changliang Zou

Asymptotic Statistics-III. Changliang Zou Asymptotic Statistics-III Changliang Zou The multivariate central limit theorem Theorem (Multivariate CLT for iid case) Let X i be iid random p-vectors with mean µ and and covariance matrix Σ. Then n (

More information

Testing Error Correction in Panel data

Testing Error Correction in Panel data University of Vienna, Dept. of Economics Master in Economics Vienna 2010 The Model (1) Westerlund (2007) consider the following DGP: y it = φ 1i + φ 2i t + z it (1) x it = x it 1 + υ it (2) where the stochastic

More information

Inference for Parameters Defined by Moment Inequalities: A Recommended Moment Selection Procedure

Inference for Parameters Defined by Moment Inequalities: A Recommended Moment Selection Procedure Inference for Parameters Defined by Moment Inequalities: A Recommended Moment Selection Procedure Donald W. K. Andrews Cowles Foundation for Research in Economics Yale University Panle Jia Department of

More information

SOLUTIONS Problem Set 2: Static Entry Games

SOLUTIONS Problem Set 2: Static Entry Games SOLUTIONS Problem Set 2: Static Entry Games Matt Grennan January 29, 2008 These are my attempt at the second problem set for the second year Ph.D. IO course at NYU with Heski Bar-Isaac and Allan Collard-Wexler

More information

Identification of Panel Data Models with Endogenous Censoring

Identification of Panel Data Models with Endogenous Censoring Identification of Panel Data Models with Endogenous Censoring S. Khan Department of Economics Duke University Durham, NC - USA shakeebk@duke.edu M. Ponomareva Department of Economics Northern Illinois

More information

Testing Many Moment Inequalities

Testing Many Moment Inequalities Testing Many Moment Inequalities Victor Chernozhukov Denis Chetverikov Kengo Kato The Institute for Fiscal Studies Department of Economics, UCL cemmap working paper CWP52/14 TESTING MANY MOMENT INEQUALITIES

More information

MCMC CONFIDENCE SETS FOR IDENTIFIED SETS. Xiaohong Chen, Timothy M. Christensen, and Elie Tamer. May 2016 COWLES FOUNDATION DISCUSSION PAPER NO.

MCMC CONFIDENCE SETS FOR IDENTIFIED SETS. Xiaohong Chen, Timothy M. Christensen, and Elie Tamer. May 2016 COWLES FOUNDATION DISCUSSION PAPER NO. MCMC CONFIDENCE SETS FOR IDENTIFIED SETS By Xiaohong Chen, Timothy M. Christensen, and Elie Tamer May 2016 COWLES FOUNDATION DISCUSSION PAPER NO. 2037 COWLES FOUNDATION FOR RESEARCH IN ECONOMICS YALE UNIVERSITY

More information

Resampling-Based Control of the FDR

Resampling-Based Control of the FDR Resampling-Based Control of the FDR Joseph P. Romano 1 Azeem S. Shaikh 2 and Michael Wolf 3 1 Departments of Economics and Statistics Stanford University 2 Department of Economics University of Chicago

More information

Multiscale Adaptive Inference on Conditional Moment Inequalities

Multiscale Adaptive Inference on Conditional Moment Inequalities Multiscale Adaptive Inference on Conditional Moment Inequalities Timothy B. Armstrong 1 Hock Peng Chan 2 1 Yale University 2 National University of Singapore June 2013 Conditional moment inequality models

More information

Statistics 612: L p spaces, metrics on spaces of probabilites, and connections to estimation

Statistics 612: L p spaces, metrics on spaces of probabilites, and connections to estimation Statistics 62: L p spaces, metrics on spaces of probabilites, and connections to estimation Moulinath Banerjee December 6, 2006 L p spaces and Hilbert spaces We first formally define L p spaces. Consider

More information

ECON 4160, Autumn term Lecture 1

ECON 4160, Autumn term Lecture 1 ECON 4160, Autumn term 2017. Lecture 1 a) Maximum Likelihood based inference. b) The bivariate normal model Ragnar Nymoen University of Oslo 24 August 2017 1 / 54 Principles of inference I Ordinary least

More information

Flexible Estimation of Treatment Effect Parameters

Flexible Estimation of Treatment Effect Parameters Flexible Estimation of Treatment Effect Parameters Thomas MaCurdy a and Xiaohong Chen b and Han Hong c Introduction Many empirical studies of program evaluations are complicated by the presence of both

More information

Lecture 13: Subsampling vs Bootstrap. Dimitris N. Politis, Joseph P. Romano, Michael Wolf

Lecture 13: Subsampling vs Bootstrap. Dimitris N. Politis, Joseph P. Romano, Michael Wolf Lecture 13: 2011 Bootstrap ) R n x n, θ P)) = τ n ˆθn θ P) Example: ˆθn = X n, τ n = n, θ = EX = µ P) ˆθ = min X n, τ n = n, θ P) = sup{x : F x) 0} ) Define: J n P), the distribution of τ n ˆθ n θ P) under

More information

Testing Hypothesis. Maura Mezzetti. Department of Economics and Finance Università Tor Vergata

Testing Hypothesis. Maura Mezzetti. Department of Economics and Finance Università Tor Vergata Maura Department of Economics and Finance Università Tor Vergata Hypothesis Testing Outline It is a mistake to confound strangeness with mystery Sherlock Holmes A Study in Scarlet Outline 1 The Power Function

More information

Chapter 7. Hypothesis Testing

Chapter 7. Hypothesis Testing Chapter 7. Hypothesis Testing Joonpyo Kim June 24, 2017 Joonpyo Kim Ch7 June 24, 2017 1 / 63 Basic Concepts of Testing Suppose that our interest centers on a random variable X which has density function

More information

What s New in Econometrics. Lecture 1

What s New in Econometrics. Lecture 1 What s New in Econometrics Lecture 1 Estimation of Average Treatment Effects Under Unconfoundedness Guido Imbens NBER Summer Institute, 2007 Outline 1. Introduction 2. Potential Outcomes 3. Estimands and

More information

Chapter 4: Asymptotic Properties of the MLE

Chapter 4: Asymptotic Properties of the MLE Chapter 4: Asymptotic Properties of the MLE Daniel O. Scharfstein 09/19/13 1 / 1 Maximum Likelihood Maximum likelihood is the most powerful tool for estimation. In this part of the course, we will consider

More information

On the Asymptotic Optimality of Empirical Likelihood for Testing Moment Restrictions

On the Asymptotic Optimality of Empirical Likelihood for Testing Moment Restrictions On the Asymptotic Optimality of Empirical Likelihood for Testing Moment Restrictions Yuichi Kitamura Department of Economics Yale University yuichi.kitamura@yale.edu Andres Santos Department of Economics

More information

A Simple Way to Calculate Confidence Intervals for Partially Identified Parameters. By Tiemen Woutersen. Draft, September

A Simple Way to Calculate Confidence Intervals for Partially Identified Parameters. By Tiemen Woutersen. Draft, September A Simple Way to Calculate Confidence Intervals for Partially Identified Parameters By Tiemen Woutersen Draft, September 006 Abstract. This note proposes a new way to calculate confidence intervals of partially

More information

arxiv: v3 [stat.me] 26 Sep 2017

arxiv: v3 [stat.me] 26 Sep 2017 Monte Carlo Confidence Sets for Identified Sets Xiaohong Chen Timothy M. Christensen Elie Tamer arxiv:165.499v3 [stat.me] 26 Sep 217 First draft: August 215; Revised September 217 Abstract In complicated/nonlinear

More information

EXPLICIT NONPARAMETRIC CONFIDENCE INTERVALS FOR THE VARIANCE WITH GUARANTEED COVERAGE

EXPLICIT NONPARAMETRIC CONFIDENCE INTERVALS FOR THE VARIANCE WITH GUARANTEED COVERAGE EXPLICIT NONPARAMETRIC CONFIDENCE INTERVALS FOR THE VARIANCE WITH GUARANTEED COVERAGE Joseph P. Romano Department of Statistics Stanford University Stanford, California 94305 romano@stat.stanford.edu Michael

More information

Confidence intervals for projections of partially identified parameters

Confidence intervals for projections of partially identified parameters Confidence intervals for projections of partially identified parameters Hiroaki Kaido Francesca Molinari Jörg Stoye The Institute for Fiscal Studies Department of Economics, UCL cemmap working paper CWP02/16

More information

THE LIMIT OF FINITE-SAMPLE SIZE AND A PROBLEM WITH SUBSAMPLING. Donald W. K. Andrews and Patrik Guggenberger. March 2007

THE LIMIT OF FINITE-SAMPLE SIZE AND A PROBLEM WITH SUBSAMPLING. Donald W. K. Andrews and Patrik Guggenberger. March 2007 THE LIMIT OF FINITE-SAMPLE SIZE AND A PROBLEM WITH SUBSAMPLING By Donald W. K. Andrews and Patrik Guggenberger March 2007 COWLES FOUNDATION DISCUSSION PAPER NO. 1605 COWLES FOUNDATION FOR RESEARCH IN ECONOMICS

More information

Identification of Nonparametric Simultaneous Equations Models with a Residual Index Structure

Identification of Nonparametric Simultaneous Equations Models with a Residual Index Structure Identification of Nonparametric Simultaneous Equations Models with a Residual Index Structure Steven T. Berry Yale University Department of Economics Cowles Foundation and NBER Philip A. Haile Yale University

More information

Notes on Blackwell s Comparison of Experiments Tilman Börgers, June 29, 2009

Notes on Blackwell s Comparison of Experiments Tilman Börgers, June 29, 2009 Notes on Blackwell s Comparison of Experiments Tilman Börgers, June 29, 2009 These notes are based on Chapter 12 of David Blackwell and M. A.Girshick, Theory of Games and Statistical Decisions, John Wiley

More information

ECO Class 6 Nonparametric Econometrics

ECO Class 6 Nonparametric Econometrics ECO 523 - Class 6 Nonparametric Econometrics Carolina Caetano Contents 1 Nonparametric instrumental variable regression 1 2 Nonparametric Estimation of Average Treatment Effects 3 2.1 Asymptotic results................................

More information

Consonance and the Closure Method in Multiple Testing. Institute for Empirical Research in Economics University of Zurich

Consonance and the Closure Method in Multiple Testing. Institute for Empirical Research in Economics University of Zurich Institute for Empirical Research in Economics University of Zurich Working Paper Series ISSN 1424-0459 Working Paper No. 446 Consonance and the Closure Method in Multiple Testing Joseph P. Romano, Azeem

More information

So far our focus has been on estimation of the parameter vector β in the. y = Xβ + u

So far our focus has been on estimation of the parameter vector β in the. y = Xβ + u Interval estimation and hypothesis tests So far our focus has been on estimation of the parameter vector β in the linear model y i = β 1 x 1i + β 2 x 2i +... + β K x Ki + u i = x iβ + u i for i = 1, 2,...,

More information

Lecture 8: Information Theory and Statistics

Lecture 8: Information Theory and Statistics Lecture 8: Information Theory and Statistics Part II: Hypothesis Testing and I-Hsiang Wang Department of Electrical Engineering National Taiwan University ihwang@ntu.edu.tw December 23, 2015 1 / 50 I-Hsiang

More information

More on Confidence Intervals for Partially Identified Parameters

More on Confidence Intervals for Partially Identified Parameters More on Confidence Intervals for Partially Identified Parameters Jörg Stoye New York University August 6, 2007 Abstract This paper extends Imbens and Manski s (2004) analysis of confidence intervals for

More information

Lecture 1: Entropy, convexity, and matrix scaling CSE 599S: Entropy optimality, Winter 2016 Instructor: James R. Lee Last updated: January 24, 2016

Lecture 1: Entropy, convexity, and matrix scaling CSE 599S: Entropy optimality, Winter 2016 Instructor: James R. Lee Last updated: January 24, 2016 Lecture 1: Entropy, convexity, and matrix scaling CSE 599S: Entropy optimality, Winter 2016 Instructor: James R. Lee Last updated: January 24, 2016 1 Entropy Since this course is about entropy maximization,

More information

Specification Test for Instrumental Variables Regression with Many Instruments

Specification Test for Instrumental Variables Regression with Many Instruments Specification Test for Instrumental Variables Regression with Many Instruments Yoonseok Lee and Ryo Okui April 009 Preliminary; comments are welcome Abstract This paper considers specification testing

More information

ASYMPTOTIC SIZE AND A PROBLEM WITH SUBSAMPLING AND WITH THE M OUT OF N BOOTSTRAP. DONALD W.K. ANDREWS and PATRIK GUGGENBERGER

ASYMPTOTIC SIZE AND A PROBLEM WITH SUBSAMPLING AND WITH THE M OUT OF N BOOTSTRAP. DONALD W.K. ANDREWS and PATRIK GUGGENBERGER ASYMPTOTIC SIZE AND A PROBLEM WITH SUBSAMPLING AND WITH THE M OUT OF N BOOTSTRAP BY DONALD W.K. ANDREWS and PATRIK GUGGENBERGER COWLES FOUNDATION PAPER NO. 1295 COWLES FOUNDATION FOR RESEARCH IN ECONOMICS

More information

Inference under shape restrictions

Inference under shape restrictions Inference under shape restrictions Joachim Freyberger Brandon Reeves June 7, 2018 Abstract We propose a uniformly valid inference method for an unknown function or parameter vector satisfying certain shape

More information

Generated Covariates in Nonparametric Estimation: A Short Review.

Generated Covariates in Nonparametric Estimation: A Short Review. Generated Covariates in Nonparametric Estimation: A Short Review. Enno Mammen, Christoph Rothe, and Melanie Schienle Abstract In many applications, covariates are not observed but have to be estimated

More information

ECON 3150/4150, Spring term Lecture 6

ECON 3150/4150, Spring term Lecture 6 ECON 3150/4150, Spring term 2013. Lecture 6 Review of theoretical statistics for econometric modelling (II) Ragnar Nymoen University of Oslo 31 January 2013 1 / 25 References to Lecture 3 and 6 Lecture

More information

Inference For High Dimensional M-estimates: Fixed Design Results

Inference For High Dimensional M-estimates: Fixed Design Results Inference For High Dimensional M-estimates: Fixed Design Results Lihua Lei, Peter Bickel and Noureddine El Karoui Department of Statistics, UC Berkeley Berkeley-Stanford Econometrics Jamboree, 2017 1/49

More information

Robust Inference for Differentiated Product Demand Systems

Robust Inference for Differentiated Product Demand Systems Robust Inference for Differentiated Product Demand Systems Xian Li Job Market Paper - Latest Version November 2, 207 Abstract This paper provides robust inference for differentiated product demand with

More information

A Simple, Graphical Procedure for Comparing Multiple Treatment Effects

A Simple, Graphical Procedure for Comparing Multiple Treatment Effects A Simple, Graphical Procedure for Comparing Multiple Treatment Effects Brennan S. Thompson and Matthew D. Webb May 15, 2015 > Abstract In this paper, we utilize a new graphical

More information

Testing Statistical Hypotheses

Testing Statistical Hypotheses E.L. Lehmann Joseph P. Romano Testing Statistical Hypotheses Third Edition 4y Springer Preface vii I Small-Sample Theory 1 1 The General Decision Problem 3 1.1 Statistical Inference and Statistical Decisions

More information

Chapter 6. Order Statistics and Quantiles. 6.1 Extreme Order Statistics

Chapter 6. Order Statistics and Quantiles. 6.1 Extreme Order Statistics Chapter 6 Order Statistics and Quantiles 61 Extreme Order Statistics Suppose we have a finite sample X 1,, X n Conditional on this sample, we define the values X 1),, X n) to be a permutation of X 1,,

More information

IDENTIFICATION AND ESTIMATION OF PARTIALLY IDENTIFIED MODELS WITH APPLICATIONS TO INDUSTRIAL ORGANIZATION

IDENTIFICATION AND ESTIMATION OF PARTIALLY IDENTIFIED MODELS WITH APPLICATIONS TO INDUSTRIAL ORGANIZATION The Pennsylvania State University The Graduate School Department of Economics IDENTIFICATION AND ESTIMATION OF PARTIALLY IDENTIFIED MODELS WITH APPLICATIONS TO INDUSTRIAL ORGANIZATION A Dissertation in

More information

Rejoinder on: Control of the false discovery rate under dependence using the bootstrap and subsampling

Rejoinder on: Control of the false discovery rate under dependence using the bootstrap and subsampling Test (2008) 17: 461 471 DOI 10.1007/s11749-008-0134-6 DISCUSSION Rejoinder on: Control of the false discovery rate under dependence using the bootstrap and subsampling Joseph P. Romano Azeem M. Shaikh

More information

Oblivious Equilibrium: A Mean Field Approximation for Large-Scale Dynamic Games

Oblivious Equilibrium: A Mean Field Approximation for Large-Scale Dynamic Games Oblivious Equilibrium: A Mean Field Approximation for Large-Scale Dynamic Games Gabriel Y. Weintraub, Lanier Benkard, and Benjamin Van Roy Stanford University {gweintra,lanierb,bvr}@stanford.edu Abstract

More information

Instrumental Variables Estimation and Weak-Identification-Robust. Inference Based on a Conditional Quantile Restriction

Instrumental Variables Estimation and Weak-Identification-Robust. Inference Based on a Conditional Quantile Restriction Instrumental Variables Estimation and Weak-Identification-Robust Inference Based on a Conditional Quantile Restriction Vadim Marmer Department of Economics University of British Columbia vadim.marmer@gmail.com

More information

A Bootstrap Test for Conditional Symmetry

A Bootstrap Test for Conditional Symmetry ANNALS OF ECONOMICS AND FINANCE 6, 51 61 005) A Bootstrap Test for Conditional Symmetry Liangjun Su Guanghua School of Management, Peking University E-mail: lsu@gsm.pku.edu.cn and Sainan Jin Guanghua School

More information

Identification and shape restrictions in nonparametric instrumental variables estimation

Identification and shape restrictions in nonparametric instrumental variables estimation Identification shape restrictions in nonparametric instrumental variables estimation Joachim Freyberger Joel Horowitz The Institute for Fiscal Studies Department of Economics, UCL cemmap woring paper CWP3/3

More information

Testing Problems with Sub-Learning Sample Complexity

Testing Problems with Sub-Learning Sample Complexity Testing Problems with Sub-Learning Sample Complexity Michael Kearns AT&T Labs Research 180 Park Avenue Florham Park, NJ, 07932 mkearns@researchattcom Dana Ron Laboratory for Computer Science, MIT 545 Technology

More information

STAT 200C: High-dimensional Statistics

STAT 200C: High-dimensional Statistics STAT 200C: High-dimensional Statistics Arash A. Amini May 30, 2018 1 / 57 Table of Contents 1 Sparse linear models Basis Pursuit and restricted null space property Sufficient conditions for RNS 2 / 57

More information

STAT 135 Lab 5 Bootstrapping and Hypothesis Testing

STAT 135 Lab 5 Bootstrapping and Hypothesis Testing STAT 135 Lab 5 Bootstrapping and Hypothesis Testing Rebecca Barter March 2, 2015 The Bootstrap Bootstrap Suppose that we are interested in estimating a parameter θ from some population with members x 1,...,

More information

Cherry Blossom run (1) The credit union Cherry Blossom Run is a 10 mile race that takes place every year in D.C. In 2009 there were participants

Cherry Blossom run (1) The credit union Cherry Blossom Run is a 10 mile race that takes place every year in D.C. In 2009 there were participants 18.650 Statistics for Applications Chapter 5: Parametric hypothesis testing 1/37 Cherry Blossom run (1) The credit union Cherry Blossom Run is a 10 mile race that takes place every year in D.C. In 2009

More information